Generating Robust Adversarial Images in Medical Image Classification using GANs

Rapid advancements in artificial intelligence and deep learning technologies have greatly improved the accurate classification of medical images, resulting in earlier diagnosis of disease. However, these advancements have also brought attention to the challenge of ensuring the resilience of models against ad- versarial medical images. In this study, we propose an adversarial attack algorithm that utilizes generative adversarial networks for medical image classification. This algorithm seamlessly integrates the target classifier into the attack generation framework, en- abling it to learn the distribution of adversarial examples closely resembling real data. Through extensive experiments across var- ious medical image classification datasets including fundoscopy, chest x-ray, skin lesion, and breast histology, the proposed attack algorithm achieves impressive attack success rates. Specifically, rates of 81.36%, 95.23%, 77.75%, and 51.67% are reported for each dataset, respectively. These findings underscore the critical importance of addressing adversarial vulnerabilities in medical image classification, ensuring the reliability and trustworthiness of diagnostic systems in clinical settings.